AC_TYPE_LONG_LONG_INT and AC_TYPE_UNSIGNED_LONG_LONG_INT (was: Re: [Bug-tar] GNU tar 1.19 on HP-UX)

[Bruno Haible]

Hi Paul,

Paul Eggert wrote on 2007-10-16:
> Thanks for verifying this.  I have started by installing the following
> fix to Autoconf, and will follow up on gnulib shortly.
> 
> 2007-10-16  Paul Eggert  <address@hidden>
> 
>       Check for 64-bit int errors in HP-UX 10.20 preprocessor.

It exists also in the HP-UX 11.00 preprocessor.

>       Problem reported by H.Merijn Brand in
>       <http://lists.gnu.org/archive/html/bug-tar/2007-10/msg00018.html>.
>       * lib/autoconf/types.m4 (AC_TYPE_LONG_LONG_INT):
>       (AC_TYPE_UNSIGNED_LONG_LONG_INT):
>       Check that preprocessor handles 64-bit ints, too.

The effect is that one of these tests says yes, and the other says no.
Using the stdint.h from gnulib, one then has this:

$ cat foo.c
#include "config.h"
#include <stdio.h>
#include <stdint.h>
uint64_t uu;
int64_t ii;
int main ()
{
  fprintf(stderr, "intmax_t size = %d\n", sizeof (intmax_t));
  fprintf(stderr, "uintmax_t size = %d\n", sizeof (uintmax_t));
  fprintf(stderr, "int64_t size = %d\n", sizeof (int64_t));
  fprintf(stderr, "uint64_t size = %d\n", sizeof (uint64_t));
}
$ cc -Ae -O -I.. -I. foo.c libgnu.a 
$ ./a.out 
intmax_t size = 8
uintmax_t size = 4
int64_t size = 8
uint64_t size = 8

Three things are not good here:
1) sizeof (uintmax_t) < sizeof (intmax_t).
2) sizeof (uintmax_t) < sizeof (uint64_t).
3) sizeof (uintmax_t) < 8.

1) is bad because much code assumes that every signed int can be losslessly
   casted into its corresponding unsigned counterpart. Also a lot of code
   assumes that sizeof (uintmax_t) == sizeof (intmax_t).
   One such code is gnulib's vasnprintf: it assumes that %jd and %ju take
   as argument an integer of the same size. In this situation it does not do
   so any more, and the unit test of the 'fprintf-posix' module fails.

2) is bad because a lot code assumes that every unsigned scalar integer type
   can be losslessly casted to uintmax_t.

3) is bad because ISO C 99 requires the existence of an uint_least64_t type;
   thus 64-bit types should be commonplace.

So the change fixed one problem but opened up three other problems.

What are the options?

A) Accept the fact that although 64-bit types may be known to the compiler,
   they may not portably be used in preprocessor expressions.
   This means:
   - In the autoconf tests: undo the last change.
   - In gnulib's stdint.h and inttypes.h: Compare the values of UINTMAX_MAX
     etc. at configure time, rather than at compile time in the preprocessor.

B) Turn off the 64-bit types of HP-UX 11 cc.
   This means:
   - Change AC_TYPE_LONG_LONG_INT so that it says "no" whenever
     AC_TYPE_UNSIGNED_LONG_LONG_INT says "no".
   - Add a "#define int64_t unusable_int64_t"
     and   "#define uint64_t unusable_uint64_t"
     to config.h, also as part of AC_TYPE_LONG_LONG_INT and
     AC_TYPE_UNSIGNED_LONG_LONG_INT.

C) Some other option I'm missing?

Bruno